hexagon: HMX quantized matmul rework

ggml/src/ggml-hexagon/htp/hmx-ops.h

@@ -33,28 +33,27 @@ typedef struct {
     size_t        src1_nb3;
     size_t        dst_nb2;
     size_t        dst_nb3;
-} hmx_matmul_w16a32_batched_params_t;
+} hmx_matmul_f16_f32_batched_params_t;
 
 // HMX matrix multiplication — tile-permuted FP16 weights, FP32 activation/output
 // act_stride: activation row stride in elements (= k for contiguous, or
 //             nb[1]/sizeof(float) for permuted tensors like attention Q).
 // weight_stride: weight row stride in elements (= k for compact weights, or
 //                nb[1]/sizeof(__fp16) for permuted KV-cache views used by QK).
-int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx,
+int hmx_matmul_f16_f32(struct htp_context *ctx,
                                 float *restrict dst,
                                 const float *activation,
                                 const __fp16 *permuted_weight,
                                 int m, int k, int n,
                                 int act_stride,
                                 int weight_stride);
 
-// Batched F16 wrapper over hmx_mat_mul_permuted_w16a32.
+// Batched F16 wrapper over hmx_mat_mul_f16_f32.
 // Batch semantics match ggml_mul_mat(): src0 broadcasts to src1 in dims 2/3.
-int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx,
-                                        const hmx_matmul_w16a32_batched_params_t *params);
+int hmx_matmul_f16_f32_batched(struct htp_context *ctx, const hmx_matmul_f16_f32_batched_params_t *params);
 
-// HMX matrix multiplication — tile-permuted quantised weights (Q4_0/Q8_0/IQ4_NL)
-int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx,
+// HMX matrix multiplication — quantised weights (Q4_0/Q8_0/IQ4_NL/MXFP4)
+int hmx_matmul_q_f32(struct htp_context *ctx,
                                       float *restrict dst,
                                       const float *activation,
                                       const uint8_t *permuted_weight,

ggml/src/ggml-hexagon/htp/main.c

@@ -87,35 +87,37 @@ AEEResult htp_iface_open(const char * uri, remote_handle64 * handle) {
         }
     }
 
+#if __HVX_ARCH__ >= 75
     {
-        // Power on HMX
+        // Power on HMX and set HMX clock
         HAP_power_request_t request;
         memset(&request, 0, sizeof(HAP_power_request_t));
-        request.type         = HAP_power_set_HMX;
-        request.hmx.power_up = TRUE;
-        FARF(ALWAYS, "Powering HMX on\n");
-        err = HAP_power_set((void *) &ctx, &request);
+        request.type = HAP_power_set_HMX_v2;
+        request.hmx_v2.set_power     = TRUE;
+        request.hmx_v2.power_up      = TRUE;
+        request.hmx_v2.set_clock     = TRUE;
+        request.hmx_v2.target_corner = HAP_DCVS_EXP_VCORNER_MAX;
+        request.hmx_v2.min_corner    = HAP_DCVS_EXP_VCORNER_MAX;
+        request.hmx_v2.max_corner    = HAP_DCVS_EXP_VCORNER_MAX;
+        request.hmx_v2.perf_mode     = HAP_CLK_PERF_HIGH;
+        FARF(ALWAYS, "Setting HMX clock\n");
+        err = HAP_power_set((void *) ctx, &request);
         if (err != AEE_SUCCESS) {
-            FARF(ERROR, "Error powering on HMX.");
+            FARF(ERROR, "Error setting HMX clock.");
             return err;
         }
     }
-
-#if __HVX_ARCH__ >= 75
+#else
     {
-        // Set HMX clock
+        // Power on HMX
         HAP_power_request_t request;
         memset(&request, 0, sizeof(HAP_power_request_t));
-        request.type = HAP_power_set_HMX_v2;
-        request.hmx_v2.set_clock = TRUE;
-        request.hmx_v2.target_corner = HAP_DCVS_EXP_VCORNER_MAX;
-        request.hmx_v2.min_corner = HAP_DCVS_EXP_VCORNER_MAX;
-        request.hmx_v2.max_corner = HAP_DCVS_EXP_VCORNER_MAX;
-        request.hmx_v2.perf_mode = HAP_CLK_PERF_HIGH;
-        FARF(ALWAYS, "Setting HMX clock\n");
-        err = HAP_power_set((void *) &ctx, &request);
+        request.type         = HAP_power_set_HMX;
+        request.hmx.power_up = TRUE;
+        FARF(ALWAYS, "Powering HMX on\n");
+        err = HAP_power_set((void *) ctx, &request);
         if (err != AEE_SUCCESS) {
-            FARF(ERROR, "Error setting HMX clock.");
+            FARF(ERROR, "Error powering on HMX.");
             return err;
         }
     }

ggml/src/ggml-hexagon/htp/matmul-ops.c

@@ -2995,7 +2995,6 @@ int op_matmul(struct htp_ops_context * octx) {
     //  is handled by HMX itself; when M < 32  fall back to HVX.
     const int m_total = (int) src1->ne[1];
     const int m_hmx   = m_total & ~31;   // 0 when M < 32
-
     if (m_hmx == 0) {
         return op_matmul_hvx(octx);
     }
@@ -3020,7 +3019,7 @@ int op_matmul(struct htp_ops_context * octx) {
 
     if (src0->type == HTP_TYPE_F16) {
         if (is_batched) {
-            hmx_matmul_w16a32_batched_params_t batch_params = {
+            hmx_matmul_f16_f32_batched_params_t batch_params = {
                 .dst             = (float *) dst->data,
                 .activation      = (float *) src1->data,
                 .permuted_weight = (const __fp16 *) src0->data,
@@ -3041,15 +3040,14 @@ int op_matmul(struct htp_ops_context * octx) {
                 .dst_nb2         = dst->nb[2],
                 .dst_nb3         = dst->nb[3],
             };
-            ret = hmx_mat_mul_permuted_w16a32_batched(octx->ctx, &batch_params);
+            ret = hmx_matmul_f16_f32_batched(octx->ctx, &batch_params);
         } else {
-            ret = hmx_mat_mul_permuted_w16a32(octx->ctx,
+            ret = hmx_matmul_f16_f32(octx->ctx,
                     (float*) dst->data, (float*) src1->data, (const __fp16 *) src0->data,
                     m_total, k, n, act_stride, wgt_stride);
         }
     } else {
-        ret = hmx_mat_mul_permuted_qk_0_d16a32(octx->ctx,
-                    (float*) dst->data, (float*) src1->data, (const uint8_t *) src0->data,
+        ret = hmx_matmul_q_f32(octx->ctx, (float*) dst->data, (float*) src1->data, (const uint8_t *) src0->data,
                     m_total, k, n, (int) src0->type);
     }
 

scripts/snapdragon/adb/run-bench.sh

@@ -45,5 +45,5 @@ adb $adbserial $adbhost shell " \
   ADSP_LIBRARY_PATH=$basedir/$branch/lib \
     $ndev $nhvx $opmask $verbose $profile $hb ./$branch/bin/llama-bench --device $device --mmap 0 -m $basedir/../gguf/$model \
         --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 \
-        --ubatch-size 256 -fa 1 -ngl 99 $cli_opts $@    \
+        --ubatch-size 1024 -fa 1 -ngl 99 $cli_opts $@   \
 "

scripts/snapdragon/adb/run-cli.sh

@@ -73,6 +73,6 @@ adb $adbserial $adbhost shell " \
     $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $opflt $vmem $mbuf \
       ./$branch/bin/llama-cli --no-mmap -m $basedir/../gguf/$model \
          --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1           \
-         --ctx-size 8192 --ubatch-size 256 -fa on                  \
+         --ctx-size 8192 --ubatch-size 1024 -fa on                 \
          -ngl 99 --device $device $cli_opts $@                     \
 "

scripts/snapdragon/adb/run-completion.sh

@@ -69,6 +69,6 @@ adb $adbserial $adbhost shell " \
     $verbose $sched $opmask $profile $nhvx $hmx $ndev $hb $opbatch $opqueue $opflt $vmem $mbuf \
       ./$branch/bin/llama-completion --no-mmap -m $basedir/../gguf/$model \
          --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1                  \
-         --ctx-size 8192 --ubatch-size 256 -fa on                         \
-         -ngl 99 --device $device $cli_opts $@                    \
+         --ctx-size 8192 --ubatch-size 1024 -fa on                        \
+         -ngl 99 --device $device $cli_opts $@                            \
 "

scripts/snapdragon/adb/run-mtmd.sh

@@ -66,6 +66,6 @@ adb $adbserial $adbhost shell " \
          --mmproj $basedir/../gguf/$mmproj                                   \
          --image $basedir/../gguf/$image                                     \
          --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1                     \
-         --ctx-size 8192 --ubatch-size 256 -fa on                            \
+         --ctx-size 8192 --ubatch-size 1024 -fa on                           \
          -ngl 99 --device $device -v $cli_opts $@                            \
 "

scripts/snapdragon/windows/run-bench.ps1

@@ -45,4 +45,4 @@ $env:ADSP_LIBRARY_PATH="$basedir\lib"
 & "$basedir\bin\llama-bench.exe" `
     --mmap 0 -m $basedir\..\..\gguf\$model `
     --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
-    --batch-size 128 -ngl 99 --device $device $cli_opts
+    --ubatch-size 1024 -ngl 99 --device $device $cli_opts

scripts/snapdragon/windows/run-cli.ps1

@@ -49,5 +49,5 @@ $env:ADSP_LIBRARY_PATH="$basedir\lib"
 & "$basedir\bin\llama-cli.exe" `
     --no-mmap -m $basedir\..\..\gguf\$model `
     --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
-    --ctx-size 8192 --ubatch-size 256 -fa on `
+    --ctx-size 8192 --ubatch-size 1024 -fa on `
     -ngl 99 --device $device $cli_opts

scripts/snapdragon/windows/run-completion.ps1

@@ -49,5 +49,5 @@ $env:ADSP_LIBRARY_PATH="$basedir\lib"
 & "$basedir\bin\llama-completion.exe" `
     --no-mmap -m $basedir\..\..\gguf\$model `
     --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
-    --ctx-size 8192 --batch-size 256 -fa on `
+    --ctx-size 8192 --ubatch-size 1024 -fa on `
     -ngl 99 -no-cnv --device $device $cli_opts

scripts/snapdragon/windows/run-mtmd.ps1

@@ -64,5 +64,5 @@ $env:ADSP_LIBRARY_PATH="$basedir\lib"
     --mmproj $basedir\..\..\gguf\$mmproj `
     --image $basedir\..\..\gguf\$image `
     --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
-    --ctx-size 8192 --ubatch-size 256 -fa on `
+    --ctx-size 8192 --ubatch-size 1024 -fa on `
     -ngl 99 --device $device -v $cli_opts